memory-audit

user-invocable	false
name	memory-audit
description	Audit and review agent memory files for quality, staleness, duplicates, and gaps. Use when user says "audit memory", "check memory", "review memory files", "what's in memory", "clean up memory", "check my memories", or "find duplicate memories".

Memory Audit

Pattern: Reviewer (Google ADK) — Load → Analyze → Report → Recommend

Audits mem0 memories stored in Qdrant, detecting duplicates, contradictions, and staleness. Outputs a structured report.

USE FOR

• "audit my memory"
• "check memory files"
• "review memory"
• "what's stored in memory?"
• "clean up memory"
• "find duplicate memories"
• "are there contradictions in my memory?"
• "memory health check"

REQUIRES

• Python 3.10+
• Qdrant running on localhost:6333
• GLM_API_KEY environment variable (for contradiction detection)
• pip install -r requirements.txt

When to Use

Use this skill whenever the user wants to inspect, clean, or review their agent's stored memories. Especially useful after extended use when memory drift or duplication may occur.

Trigger keywords: audit memory, check memory, memory quality, memory cleanup, duplicate memories, stale memory.

Don't use when: You're looking for a specific memory or fact (just search directly). This skill audits quality, not retrieves content.

Prerequisites

• Qdrant vector database running: docker ps | grep qdrant
• Collection memories exists in Qdrant
• Python 3.10+ installed: python3 --version
• Dependencies installed: pip install -r ~/.openclaw/skills/memory-audit/requirements.txt
• GLM API key set: echo $GLM_API_KEY

Quick Start

# Run full memory audit
cd ~/.openclaw/skills/memory-audit
./run.sh

# View results
cat ~/.openclaw/workspace/memory-audit-report.md

Instructions

What the Audit Does

1. Pull memories — Fetch all vectors + payloads from Qdrant (localhost:6333, collection memories)
2. Group by topic — Cluster memories by keyword similarity
3. Flag duplicates — Cosine similarity > 0.85 within same topic
4. Detect contradictions — Use GLM-4.7 to compare conflicting entries
5. Generate report — Output structured Markdown to ~/.openclaw/workspace/memory-audit-report.md

Interpreting Results

Section	Meaning
## Duplicates	Memories that say essentially the same thing
## Contradictions	Memories that conflict with each other
## Topics	Grouped view of all memory entries
## Recommendations	Suggested deletions or merges

Example

User says: "check my memory for duplicates" → Steps: run ./run.sh audit script → Output: JSON/Markdown report with duplicates, contradictions, recommendations → Reply: "Found 7 duplicates and 2 contradictions. Report saved to ~/.openclaw/workspace/memory-audit-report.md. Recommend deleting 5 redundant entries."

$ ./run.sh

🔍 Fetching memories from Qdrant...
📦 Found 143 memories across 12 topics
🔁 Flagging duplicates...  → 7 duplicates found
⚡ Detecting contradictions... → 2 contradictions found
📝 Report saved to: ~/.openclaw/workspace/memory-audit-report.md

Memory Health: ⚠️  Needs attention (7 duplicates, 2 contradictions)

Sample report output:

## Duplicates (7)
- [ID: abc123] "User prefers dark mode" ↔ [ID: def456] "User likes dark themes" (similarity: 0.91)
  → Recommendation: Keep abc123, delete def456

## Contradictions (2)
- [ID: ghi789] "User works in Python" ↔ [ID: jkl012] "User primarily uses JavaScript"
  → Manual review required

Error Handling

Error	Cause	Fix
Connection refused: localhost:6333	Qdrant not running	docker start qdrant
Collection 'memories' not found	Empty or uninitialized memory	Verify agent has stored memories
GLM_API_KEY not set	Missing env var	Set in ~/.openclaw/openclaw.json
ModuleNotFoundError	Missing Python deps	pip install -r requirements.txt
Python 3.10+ required	Old Python version	Upgrade or use python3.10 explicitly

Scripts

Script	Purpose
run.sh	Main entry point — installs deps and runs audit
memory-audit.py	Core Python audit logic
requirements.txt	Python dependencies